A Probabilistic Representation of Causation: Some Implications

Patricia Cheng

The goal of causal learning is to differentiate causal relations -- event sequences on which one bases explanation and control -- from merely correlational or otherwise noncausal relations. People act as if they have answers to many related questions arising from this goal, such as: What are the conditions under which causal inference is justifiable? Why should these be the conditions? Given these conditions and a set of observations, how does one decide whether a causal relation exists? How does one decide whether multiple factors interact to produce an outcome? When there are multiple independent causes of an outcome, how does one decide how often the outcome is due to a particular cause of it? This paper illustrates that coherent answers to these questions can be formulated based on an explicit representation of the probability with which a cause influences (i.e., produces or prevents) an outcome, independently of other possible causes of the outcome. This generativity assumption differentiates causal models from associative models, including both psychological and statistical variants. Even though the generativity assumption appears to only concern causal strength, its implications go beyond parameter estimation.

Back to Probabilistic Models of Cognition: The Mathematics of Mind